First, don't blame yourself: I've been burned by the bill too

Wednesday night, 10pm. I'm at home in Hangzhou, opening my weekly report — another ¥183 tacked onto the AI bill.

The problem isn't that AI is expensive — tasks aren't sorted by weight

I know that sting. I got stuck here too: to prove I was "really using AI," I fired up the strongest, priciest model for every task. The weekly number looked great, but the actual work didn't move. Think of tokens as the unit AI charges by when processing text — using more doesn't mean more professional. A Silicon Valley 101 episode captured the turning point: Uber burned through its entire yearly AI budget in four months, Meta set usage caps for employees; Huang Dongxu also shared he used to spend $300–$400 a day. My fix splits tasks into three tiers: high-stakes work (signing contracts, client proposals) goes to frontier models; rewriting headlines, cleaning up meeting notes, drafting — cheap models or local open-weight models; passwords, ID numbers, unpublished financials — don't upload at all. Late last year, Huang Dongxu sat at his office computer, offloading many tasks to frontier models at $300–$400 daily, then shifted to a "local open-weight + frontier" combo. OpenClaw and Hermes' evolution shows this too — we're turning one-off chats into repeatable steps.

Replicate today: try it inside tools you already use, no new hardware needed

Money: reuse your existing AI tools, $0 added; if you run locally, the episode cited roughly $1/day in electricity, not counting the machine and setup. Time: 30 minutes to list your last week's 10 tasks, then pilot for 7 days. Technical barrier: routing inside chat boxes, no coding needed; installing a model on your computer requires downloads, config, and trial-and-error — not trivial. Not everyone needs this, and a local model isn't required. Skipping today is fine. First step: open your usual AI, hit "new chat," paste: "Please sort the 10 tasks I paste from my last week into A — strong model, B — cheap or local model, C — sensitive, do not submit, and tag each with estimated cost and rework time."

By stage, here's how I'd start

If just starting out with no stable income yet, I'd do the three tiers on paper or a spreadsheet first, no rush to buy gear; with 1–2 clients, I'd route the repeatable stuff first — weekly reports, headlines, FAQs — and log cost and rework for 7 days straight; if scaling up, I'd have multiple people share one table, and set a monthly token alert line. Agents will keep passing tasks to each other — costs don't disappear automatically. If the product is just a chat box with a new name, I'd first capture the client's own workflows, preferences, and materials, then write down the sensitive info, the reviewer, and the cost basis — only then talk about automation.